QingboKang/SonoReasoner-8B

SonoReasoner-8B: Ultrasound Vision-Language Model

SonoReasoner-8B is an 8 billion parameter vision-language model developed by QingboKang, specialized for anatomy-grounded ultrasound reasoning. It is built upon the Qwen/Qwen3-VL-8B-Instruct base model and undergoes a two-stage post-training process.

Key Capabilities

• Hierarchical Reasoning: Initialized using SonoCorpus, it develops reasoning capabilities across protocol, system, and organ levels.
• Clinical Task Alignment: Utilizes Group Relative Policy Optimization (GRPO) on a multi-task ultrasound mixture to align with clinical preferences.
• Multimodal Understanding: Processes both ultrasound images and text inputs.
• Specific Task Performance: Evaluated on SonoVQA and downstream tasks including diagnosis classification, lesion localization, and report generation.

Good For

• Research Use: Intended for studying hierarchical ultrasound reasoning, protocol-system-organ grounding, and ultrasound VQA.
• Medical Image Analysis: Applicable for research in multimodal reasoning within the medical imaging domain.
• Ultrasound Report Generation: Supports the generation of structured reports based on ultrasound imagery.

Limitations

• Modality Specificity: Trained exclusively on ultrasound images; generalization to other medical imaging modalities (CT, MRI, X-ray) is not guaranteed.
• Data Bias: Performance may be affected by rare findings or underrepresented protocols in the training data, and biases from SonoVQA and SonoCorpus may propagate.
• Research Only: Not intended for clinical diagnosis, treatment planning, or autonomous medical decision-making. Outputs should not be used as medical advice.